As computer systems and networks become increasingly complex and include increasing amounts of distributed or parallel processing, the method by which a system allocates usage of common resources can have a significant effect on the overall capability of a system. Such allocation methods and their supporting hardware define arbitration protocols by which the system arbitrates among agents competing for the same resource. While the problem of arbitration may arise with respect to a variety of system resources, such as processors, storage devices or memory segments, it is best characterized and understood when the resource is a communication channel, typically a communication bus.
Computer buses are used to communicate and transfer data between various parts of a computer or to or from peripheral devices. Buses may be confined to a circuit board, used to communicate from one circuit board to another, or even used to allow different processing units to communicate. This invention is useful in the arbitration of control of buses that accommodate a priority arbiter, such as a parallel contention arbiter of the type described in UK Patent Specification No 1099575, and by Taub D.M. "Arbitration and Control Acquisition in the Propose IEEE 896 Futurebus," IEEE Micro, 4:4, pp. 52-62 (1984) One particularly relevant application for such a bus arbiter is for control of buses that are used for communicating between multiple processors located in one or more computing machines.
Buses are electrical conductors, usually copper wire, copper patterns or aluminum patterns. A bus is simply the parallel electrical lines used to carry a designated number of signals with a specified format for data or addresses applied to the bus.
When an agent (a device, e.g. a processor) wishes to communicate with another agent, the first agent sends signals that cause the second agent to respond. These signals are collectively called the address or identity. The device that begins the communication is called the master, and the device that responds is called the slave. Some agents can act only as masters, some only as slaves, and other as either masters or slaves.
It is inherent in a bus that only one agent at a time may use the bus, otherwise indefinite conditions would exist. When two agents attempt to use the bus at the same time, a mechanism or protocol must decide which agent shall be able to use the bus. This mechanism is called arbitration. Details of buses and arbitration schemes may be found in Computer Buses--A Tutorial by David B. Gustavson, August 1984, IEEE Micro, p. 7, herein incorporated by reference.
The "star" arbitration scheme uses a central arbiter which is connected to each agent by two dedicated wires. One wire is used to allow the agent to send a request signal to the central arbiter, and the second wire is used to allow the central arbiter to send a bus granted signal to the agent. Thus for N agents, two * N extra wires are needed to arbitrate bus access. The central arbiter can use any of a number of schemes to arbitrate requests for the bus in a fast and efficient manner. However, such special wiring is comparatively expensive.
The so-called daisy chain method is a second arbitration scheme. Under this protocol, all agents share a bus request line and each agent has one daisy-in input line and one daisy-out output line. The daisy-in line receives an input signal from the previous agent, and the daisy-out line is connected to send an output signal to the next adjacent agent. When an agent does not want control of the bus, the signal on the daisy-in input is passed through to the daisy out output. When an agent wants control of the bus it forces a "zero" on the daisy-out output and a "one" on the shared bus request line. The shared bus request line is a wired-OR line that may be actuated by any agent. Any agent wishing to use the bus thus asserts the shared bus request line by asserting a logical "one" on the line. This places a one on the daisy-in input of the highest priority agent. If the first priority agent does not want to use the bus, it passes the one through to its daisy-out output. The signal is passed along in a daisy-chain manner until an agent seeking control of the bus receives a logical one on its daisy-in in-put. If no agent preceding the asserting agent wants the bus, the signal from the shared bus request line will be passed through to the asserting 1 agent, and a bus grant signal will appear on the asserting agent's daisy-in in-put. If a preceding agent desires the bus, the preceding agent will gain control of the bus and will continue placing a zero on its daisy-out output, and all subsequent agents (including the asserting agent) will be inhibited from using the bus. Thus, the closer an agent is to the beginning of the daisy chain, the higher priority the agent has. Also, rules are necessary to prevent a high priority agent from taking the bus away from a low priority agent while it is in use. This may be accomplished by synchronizing request assertions. The daisy chain is economical but slow, and it requires every agent be in place to complete the daisy chain bus grant line. This scheme also does not have much fairness among the agents, since higher priority agents will be able to disproportionately dominate control of the bus.
A third scheme, the parallel contention arbiter as described in the references cited above, is based on assigning each agent a unique fixed k-bit arbitration number called its "identity." The value of k is at least [log.sub.2 (N+1)], where N is the maximum number of agents that can be attached to the bus. An agent that wants control of the bus forces a wired-OR shared bus request line, and waits for a signal to start arbitration. The signal to start arbitration may be generated by the current bus master, a central timing controller, or any agent on the bus. At the start of an arbitration, the agent applies its arbitration number to a separate parallel set of arbitration control lines provided on the bus for this purpose. The agent then monitors each of the arbitration lines, in parallel. If the value carried by line i is "1," but the agent is applying a "0" to it, then the agent removes the lower-order (i-1 to 0) bits of its identity. If line i drops back to "0," the agent reapplies the lower-order bits it removed before. For example, consider the case where two agents with identities 1010101 and 0011100, respectively, are requesting the bus. The first agent will remove its three lowest order bits, leaving 1010000, and the second agent will remove all of its bits. Next, the first agent will reapply its three lowest order bits, and the second agent will do nothing, since the most significant bit still remains. It is easy to see that after some period of time the system reaches steady state, in which the lines carry the largest identity of all competing agents. The agent whose arbitration number matches the winning number becomes the next bus master. Note that at the end of the arbitration, each agent knows the identity of the winner, as well as whether it has won or lost. The parallel contention arbiter thus described is a priority arbiter that is very fast but is also inherently unfair, since an agent with a higher identity will always win an arbitration against an agent with a lower identity, and thus higher priority agents will disproportionately control the bus.
The parallel contention arbiter, due to its low cost and high speed, has been adopted in several multiprocessor system bus designs. To overcome the unfairness inherent in the basic priority selection mechanism, assured access protocols have been designed to provide a set of agents (i.e. the processors) with equal access to the bus. These protocols are based on batching requests, such that all requests in a batch are served before any new requests can be made. In particular, requests in the batch from agents with low assigned identities will receive service before new requests can be made by agents with high assigned identities.
In one protocol, all requests that arrive to an idle bus assert the bus request line and form a batch. An agent in the batch competes during each arbitration until it has been granted ownership of the bus. An agent that generates a new request while a batch is in progress must wait for the batch to end before asserting the request line and competing for access. The end of the batch is generally signalled by a logical "0" on the request line, since each agent in the batch releases the request line at the start of its bus tenure. All requests that are waiting at the end of a batch assert the shared request line and form a new batch. Agents in a batch receive service in order of their assigned identities, according to the parallel contention arbitration.
In a second assured access protocol, an agent with a request asserts the request line and competes in successive arbitrations until acquiring the bus. At the completion of its bus tenure, the agent marks itself as "inhibited," and won't assert the request line or compete for bus ownership until a fairness release operation takes place. The fairness release operation is a cycle in which no agents assert the request line. In other words, either there are no outstanding requests, or all agents with outstanding requests are inhibited.
The second protocol implements a batching algorithm similar to the first protocol. A batch starts and ends with a fairness release cycle. No agent is bus master more than once in a batch, but an agent with a request that is generated during a batch is allowed to join the batch if the agent has not previously received service in the batch.
There is a source of unfairness even in the above assured access protocols. In every batch, an agent receives service after all agents in the batch that have higher identities have received service. For example, in multiprocessor systems in which the processors do not continue executing while waiting for a memory request to be satisfied, this means that the lower-identity processors execute at a slower rate. The difference in throughput between the most favorably treated agent (i.e. the agent with the highest assigned identity) and the least favorably treated agent may be 10%, and can be as high as 100% for each of the protocols described above. Tightly coupled parallel algorithms are often sensitive to the speed of the slowest processor. In this case, the unfairness can affect total system performance.
Priority scheduling of urgent requests may be integrated with the assured access protocols in the parallel contention arbiter. In this case, agents follow the assured access protocol for non-priority requests, but ignore the protocol and compete in every arbitration for priority requests. Furthermore, an extra line can be provided on the bus, to be treated as the most significant bit of the agent's identity. Agents with priority requests assert this line during arbitration; agents with non-priority requests do not. This guarantees that all priority requests will be served before non-priority requests.
Prior art arbitration protocols have not been able to economically insure fairness. The star arbitration method may be fair, but it also may be expensive. The assured access protocols for the parallel contention arbiter are relatively inexpensive, but are unfair. The present invention successfully implements a fair arbitration protocol for priority arbiters. When used in a parallel contention arbiter, fairness can be inexpensively achieved.